Cascading failure: auto-compaction, /compact, and rewind all fail at context limit — requires Ctrl+C to escape

[sjdodge123]

Description

When a conversation hits the context limit, a cascading series of failures occurs with no graceful recovery:

1. Auto-compaction triggers when context limit is reached, but errors out:

   Error: Error during compaction: Error: Conversation too long. Press esc twice to go up a few messages and try again.
   

2. Manual /compact also fails with the same error — despite being the suggested remedy for hitting the context limit.

3. Rewind is offered as a fallback, but that also fails:

   Error: Failed to summarize:
   Error: Conversation too long. Press esc twice to go up a few messages and try again.
   

4. Session becomes unresponsive after the rewind error — the only way to escape was Ctrl+C to kill the process. No further input was accepted.

Screenshot

(See description above for the full error sequence)

Impact

Complete loss of session context and any in-progress work state. The user is forced to hard-kill the process and start fresh with /clear or a new session.

Expected Behavior

• Compaction (auto or manual) should handle conversations at the context limit — that's its primary purpose
• If compaction fails, the session should remain interactive (not hang)
• There should always be a graceful exit path that doesn't require Ctrl+C

Environment

• Claude Code CLI
• Model: claude-opus-4-6
• macOS (Darwin 25.1.0)

[github-actions]

Found 3 possible duplicate issues:

1. Deadlock: context limit reached but /compact and rewind both fail with 'Conversation too long' #23809
2. [Bug] Auto-Compact Failure Leaves Instance in Unrecoverable Context Limit State #24942
3. /compact fails with 'Conversation too long' when context limit is reached #23047

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

[ThatDragonOverThere]

Hit this exact deadlock multiple times in the last 24 hours on Windows, Opus 4.6. Auto-compact fails, /compact fails, rewind fails. The only option is Ctrl+C, which then chains into two more bugs:

1. PowerShell escape (CLI: Task execution escapes to PowerShell, becomes unresponsive #21576) — after Ctrl+C, sometimes drops to PowerShell instead of staying in Claude Code
2. Stale resume (--resume loads hours-old context instead of current session state #26246) — --resume loads context from hours ago, losing all recent work

This creates a cascade: premature context limit → deadlock → forced exit → stale resume → hours of work lost. Happening multiple times per day now.

[chickensintrees]

Same cascade, confirmed. Filed #26575 about the rate-limit trigger variant — same unrecoverable outcome, different entry point.

ThatDragonOverThere's description of the full chain (premature limit → deadlock → forced exit → stale resume → hours lost, multiple times per day) is exactly right. We were hitting this during a client deliverable sprint with three concurrent 1M sessions.

Our defensive approach: the agent writes critical state to disk continuously (not just at session end), so when a session gets bricked, we lose the conversation but not the decisions. It's duct tape, but it works. The underlying fix needs to come from Anthropic — compaction and session management should be exempt from rate limits, full stop.

Saw your groombot project — using ChatGPT Desktop to groom Linear epics before handing to Claude/Codex/Gemini. Smart pipeline thinking. We're doing something similar with agent orchestration on the Claude Code side (lead agent delegates to sub-agents to keep the main context lean). Would be curious how the multi-tool pipeline compares.

[scapeshift-ojones]

Your cascading failure scenario compounds with the issues tracked in #27242: even when compaction succeeds, there's no way to view pre-compaction history (Ctrl+O/E broken), navigate conversation branches (right-arrow dead in picker), or link back to plan-mode predecessor sessions. And when compaction fails as you describe, you're stuck with no recovery path at all.

If this is still affecting you, please 👍 and comment on #27242 to help push past the triage threshold — we're consolidating 13+ related issues there.

[joomlafun]

- Working and never had issue event with versions before : 2.1.29
- Broken: 2.1.53
- Specific behavior: auto-compact missing, manual compact fails at limit

What difference between versions could be responsible for that?